Re: [PATCH] sched, cpumask: don't leak impossible cpus via for_each_cpu_wrap().

[Yury Norov <yury.norov@gmail.com>]

On Wed, Aug 03, 2022 at 10:49:57AM -0700, Neel Natu wrote:
> On Tue, Aug 2, 2022 at 6:22 PM Yury Norov <yury.norov@gmail.com> wrote:
> >
> > On Tue, Aug 2, 2022 at 2:41 PM Neel Natu <neelnatu@google.com> wrote:
> > >
> > > The value of 'nr_cpumask_bits' is dependent on CONFIG_CPUMASK_OFFSTACK.
> > > This in turn can change the set of cpus visited by for_each_cpu_wrap()
> > > with a mask that has bits set in the range [nr_cpu_ids, NR_CPUS).
> > >
> > > Specifically on !CONFIG_CPUMASK_OFFSTACK kernels the API can iterate
> > > over cpus outside the 'cpu_possible_mask'.
> > >
> > > Fix this to make its behavior match for_each_cpu() which always limits
> > > the iteration to the range [0, nr_cpu_ids).
> > >
> > > Signed-off-by: Neel Natu <neelnatu@google.com>
> >
> > The patch itself doesn't look correct because it randomly switches a piece
> > of cpumask API from nr_cpumask_bits to nr_cpu_ids, and doesn't touch
> > others.
> >
> > However...
> >
> > I don't know the story behind having 2 variables holding the max possible
> > number of cpus, and it looks like it dates back to prehistoric times.  In
> > modern kernel, there are 2 cases where nr_cpumask_bits == nr_cpu_ids
> > for sure: it's CONFIG_CPUMASK_OFFSTACK=y and
> > CONFIG_HOTPLUG_CPU=y. At least one of those is enabled in defconfig
> > of every popular architecture.
> >
> 
> Hmm, in a kernel with CONFIG_HOTPLUG_CPU=y but not CONFIG_CPUMASK_OFFSTACK
> I see "nr_cpu_ids = 20, nr_cpumask_bits = 512". FYI since it doesn't
> match the observation
> above that nr_cpumask_bits == nr_cpu_ids when CONFIG_HOTPLUG_CPU=y.

I said this because the comment in include/linux/cpumaks.h says:
 *  If HOTPLUG is enabled, then cpu_possible_mask is forced to have
 *  all NR_CPUS bits set, otherwise it is just the set of CPUs that
 *  ACPI reports present at boot.

And because of the code that sets nr_cpu_ids:

void __init setup_nr_cpu_ids(void)
{
        nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
}

Some arches override it, so it may be an issue. Are you running x86,
or maybe you have "nr_cpus" boot parameter?

But anyways, I feel like this should be investigated for more... Can you
please share the config of your system and boot params?
 
> > In case of HOTPLUG is off, I don't understand why we should have nr_cpu_ids
> > and nr_cpumask_bits different - what case should it cover?... Interestingly, in
> > comments to cpumask functions and in the code those two are referred
> > interchangeably.
> >
> > Even more interestingly, we have a function bitmap_setall() that sets all bits
> > up to nr_cpumask_bits, and it could trigger the problem that you described,
> 
> I think you mean cpumask_setall() that in turn calls
> bitmap_fill(nr_cpumask_bits)?

Yes, sorry.

> > so that someone would complain. (Are there any other valid reasons to set
> > bits behind nr_cpu_ids intentionally?)
> >
> 
> I don't know of any although this wasn't the case that trigger in my case.
> 
> > Can you share more details about how you triggered that? If you observe
> > those bits set, something else is probably already wrong...
> 
> The non-intuitive behavior of for_each_cpu_wrap() was triggered when iterating
> over a cpumask passed by userspace that set a bit in the [nr_cpu_ids,
> nr_cpumask_bits)
> range.

If you receive bitmap from userspace, you need to copy it with
bitmap_copy_clear_tail(), or bitmap_from_arr{32,64}, as appropriate.
That will clear unneeded bits.

For user-to-kernel communications, I'd recommend passing bitmaps in a
human-readable formats - hex string or bitmap list. Check the functions
cpumask_parse_user() and cpumask_parselist_user(). This would help to
avoid all possible issues related to endianness and 32/64 word length
mismatch.

> The kernel code is out of tree but open source so happy to provide a
> pointer if needed.

Yes please

> I considered ANDing the user supplied mask with 'cpu_possible_mask'
> but that felt
> like working around an inconsistency in the kernel API (hence the proposed fix).
> 
> > So, if there is no real case in modern kernel to have nr_cpumask_bits and
> > nr_cpu_ids different, the proper fix would be to just drop the first.
> >
> 
> I'll let other people more knowledgable than me in this area chime in.
> I'll be happy either
> way if that fixes the problem at hand.
> 
> best
> Neel
> 
> > If there is such a case, this is probably your case, and we'd know more
> > details to understand how to deal with that.
> >
> > Thanks,
> > Yury